Beyond the bot: Mapping the agentic trading stack
TL;DR: Agentic trading puts agents to work researching markets, building strategies, and placing trades within user-set limits. As agents themselves become a commodity, we believe lasting value will accrue to businesses that turn users’ goals into ongoing investment strategies, and to the platforms users trust to hold their assets. To support this transition, Coinbase Ventures is investing across three key pillars: agentic trading harnesses and terminals, access to a real edge, and compliance and observability infrastructure.

Coinbase Ventures portfolio companies are denoted with an asterisk (*) when first referenced in the article below.
AI agents are becoming active participants in the economy. Their ability to pursue goals, make decisions, and transact on our behalf is laying the foundation for an entirely new financial system: Agentic Finance, or AiFi. It runs along two lanes, both of which are changing how we interact with money. The first lane is payments and commerce, where agents pay for goods, services, and data. The second lane involves trading and markets, where agents take part directly in financial markets. This post focuses on the second lane: agentic trading.
In agentic trading, an AI agent runs part or all of the trading decision loop on your behalf. You set a goal and guardrails, and the agent can research the market, build a strategy, size positions, place trades, and monitor risk within the authority you’ve delegated. While early, this approach could make complex trading workflows accessible through simple prompts.
In our 2024 thesis on Crypto x AI, we predicted that natural-language interfaces would become the primary way people transact onchain. In our May 2025 piece, The Rise of Onchain AI, we explored how agents could make personal portfolio management more accessible. Since then, that idea has grown well beyond crypto. Exchanges, brokerages, and a new wave of startups are now racing to build ways for agents to trade across crypto, equities, derivatives, prediction markets, and more.
As the agents themselves become easier to build, we’re focused on what gives the businesses behind them an advantage that lasts. We believe value will move beyond simple order execution toward whoever manages the ongoing investment strategy and whoever users and their agents trust to hold their assets.
That view informs three areas where Coinbase Ventures is looking to invest: agentic trading harnesses and terminals, access to a real edge through specialized intelligence and data, and compliance and observability infrastructure that makes agents’ actions traceable and accountable. Below, we map the emerging landscape and explain what we believe it takes to build a durable business within it.
What is agentic trading?
Agentic trading is when an AI agent runs part or all of the trading decision loop under authority you've delegated to it. It holds a persistent objective, gathers its own information, forms and revises a strategy, and executes trades within the limits you set, without needing you to approve every step.
Trading has been automated for decades, but every prior wave of automation took a decision a human had already made and carried it out faster or cheaper:
Rule-based trading bots fire a condition you wrote in advance ("buy when X crosses Y"). They're a faster trigger, not a decision-maker.
Algorithmic execution (like TWAP or VWAP) figures out how to fill an order you already chose, never what to buy.
Robo-advisors manage portfolios using fixed rules. They can't form a new view or react to the news.
Copy trading mirrors someone else's trades, inheriting their risk without adapting to yours.
Agentic trading flips this model. You state an objective in plain language, and the agent breaks it down into research, strategy, position sizing, execution, and ongoing monitoring. For the first time, the machine forms the investment view instead of simply executing one it was given. The human moves from operator to principal: you're no longer pushing the buttons, you're setting the mandate.
Why now?
AI trading bots aren't a new idea. Many launched in 2023 and most didn't live up to the hype. We think this time is different because four things changed over the past 12 months:
AI models got reliable enough. The 2023 generation of bots failed because language models could summarize a market but couldn't execute a plan against it. Today's frontier models can handle multi-step tasks, use tools, and recover from errors, which is the minimum bar for managing positions without a human watching.
Connecting to markets got cheap. Open standards like the Model Context Protocol (MCP) and Agent2Agent (A2A) turned trading venues, data feeds, and custody systems into plug-in capabilities. Connecting an agent to a broker used to take a quarter of custom engineering. Now it can take minutes.
Open-source agent "harnesses" made it a weekend project. Tools like OpenClaw and Nous Hermes* give any model memory, scheduled tasks, and access to markets out of the box. OpenClaw alone gathered roughly 250K GitHub stars in about 60 days. Crypto absorbed this wave first because open, permissionless APIs meant no gatekeepers, which is why perpetual futures on Hyperliquid and prediction markets on Polymarket were among the earliest targets.
Markets started to converge. Tokenized stocks, perpetual futures on real-world assets, 24/7 markets, and stablecoin settlement mean a single agent can now pursue one goal across many asset classes from a single pool of collateral. Prior automation was stuck in silos because each asset class ran on separate infrastructure.
Why it matters
Agentic trading is poised to democratize the quant desk. Strategies that used to require a quantitative trading background, like blending signals across asset classes, adjusting position sizes as markets move, spotting shifts in market regime, or rebalancing with taxes in mind, are becoming something you can ask for in a sentence.
Consider a request like: "Build me a momentum-tilted portfolio across crypto and semiconductor stocks, cap my drawdown at 15%, and rebalance weekly." From that one sentence, an agent can pull momentum signals across both markets, construct the portfolio, size each position to respect the drawdown limit, place the orders across venues, and revisit the plan every week.
This model also changes what users want from a trading platform. People increasingly show up with portfolio intent ("build me a portfolio that expresses my belief in X") rather than order intent ("buy 10 shares of X"). That's a fundamentally different product than a trading screen, and it favors platforms that offer the widest range of assets over platforms that offer the best execution on a single one.
The agentic trading stack
Agentic trading is evolving from a patchwork of experimental tools into a more structured stack. We break it down into four layers: Intelligence (the data and models agents think with), Agency (the harnesses and apps that turn models into persistent trading agents), Execution (the rails that turn decisions into filled trades), and Compliance & Observability (the trust infrastructure that verifies what agents did and why).
The market map below captures the startups and platforms we're tracking across each layer, starting with intelligence at the base and building up to execution and trust at the top.

Intelligence: data and reasoning
Market data. Agents need clean, machine-readable information: prices, order books, filings, news, sentiment, and onchain activity, with timestamps and sources attached. Today's data business is built for humans, priced per seat for analysts at a terminal, not for agents making thousands of queries a day. That mismatch is an opening for machine-payable data. Coinbase's open-source x402 protocol, for example, lets an agent pay for exactly the data it needs, when it needs it, using stablecoins. Established data providers like Bloomberg, S&P Global, onchain analytics platforms like Nansen* and Dune*, and oracle networks like Pyth* and Chainlink already supply much of this information. Newer players like Surf* and Cambrian are building research and data feeds designed for agents from day one.
Market reasoning models. General-purpose reasoning from frontier labs is powerful, but it's available to everyone, which means it isn't a competitive advantage on its own. The open opportunity is purpose-built, market reasoning models: models trained specifically to reason about financial markets, test and validate a trade idea, deploy live capital and manage risk over long, unsupervised workflows. A new class of market reasoning model labs, including the likes of Prodigy Research*, EdotEnv*, and Nof1, are pursuing this, often by training on market data or building simulated trading environments for models to learn in. Alongside them, inference platforms and routers like Venice* let agents pick the best model for each task and, in several cases, pay for it per request in USDC.
Agency: harnesses, terminals, and copilots
General-purpose agent harnesses turn a model into a persistent agent by giving it memory, scheduled triggers, tool access, sandboxing, and the ability to recover from errors. They are built to handle any task, from coding to research to managing an inbox, and users extend them with plug-in skills. Open-source harnesses such as Nous Hermes* and OpenClaw have become increasingly popular in this area, and trading is quickly becoming one of their top use cases. But trading needs more than a general toolkit. We see a clear gap for "trading-hardened" harnesses with built-in skills like dynamic position sizing, take-profit and stop-loss limits, and position trimming, and a way to test a strategy before real money is on the line.
Trading agent harnesses are built for that gap. Teams like Senpi* and Supertake* are two examples. Senpi, which calls itself an "AI quant engineer," is built for active traders: it turns a trading strategy described in plain English into code and runs it around the clock across hundreds of Hyperliquid markets, applying risk-based position sizing and trailing stop-losses automatically. Supertake is built for everyday investors: users write an investment idea, or "take," in their own words, and its agent turns it into a portfolio, with a practice mode for simulated results and a public leaderboard of top-performing takes.
Agentic trading terminals are destinations. They are full workspaces where a user takes an idea through the whole loop: research, strategy design, backtesting, execution, and ongoing monitoring. Many are multi-asset, so a single agent can trade stocks, crypto, perpetual futures, and prediction markets, and most connect to existing brokerages and exchanges rather than holding assets themselves. Think of them as a trading terminal rebuilt for the agent era, aimed at active traders who want depth and control. Catalyst*, for example, turns a user's market view into a strategy that trades across stocks, options, and prediction markets, while TrueNorth gives users personalized agents that scan equities, commodities, and crypto perp markets and set up trades.
Onchain trading copilots are assistants that come to you. They are chat-first and often live inside apps people already use, such as X, Telegram, or a crypto wallet. They are built around the user's own onchain wallet, so funds usually stay self-custodied. They focus on crypto and handle everyday actions like swaps, bridges, perpetual trades, yield, and prediction market bets in a single message. Where terminals optimize for depth, copilots optimize for convenience and reach. HeyElsa*, for example, lets users build a portfolio, swap across chains, and earn yield through simple conversation, while Bankr* lets users buy, sell, and swap tokens or set limit orders by tagging it on X.
In practice, the lines between these categories are already blurring. General-purpose harnesses are adding trading skills, trading harnesses are adding the portfolios and leaderboards of a terminal, terminals are adding chat, and copilots are adding strategies that run around the clock. Over time, we expect all of them to compete on the same question: who owns the decision loop. Terminals have the edge on depth, asset breadth, and stickiness; copilots and general-purpose harnesses have the edge on distribution, since they already sit where users spend their time. The products most likely to last will pair a real edge in intelligence or strategy with distribution that is hard to copy.
Execution: venue connectivity and asset breadth
The execution layer turns an agent's decision into a filled trade. That means routing orders across venues and asset classes, tracking order status, reconciling positions, and holding the assets safely. The industry is converging on a "one credential, many assets" model, where a single agent connection can trade across spot crypto, derivatives, equities, and more.
Brokerages and exchanges are opening up to agents in two ways. The first is bring-your-own-agent access: outside agents like ChatGPT, Claude, or Grok connect through MCP servers, command-line tools, APIs, or dedicated agent accounts. Coinbase for Agents, for example, lets a third-party agent trade spot crypto, US futures, and equities on a user's Coinbase account within limits the user sets. The second is built-in intelligence: the venue's own AI works inside the app to help users research, plan, and act. Coinbase Advisor is an example. Separately, AI-native infrastructure like Alpaca and onchain trading APIs like Definitive* (Flash) give agents trading execution rails without building them from scratch.
Compliance & Observability: trust for autonomous decisions
Today's regulatory and compliance frameworks assume a human is making the call. When an agent makes the decision, someone needs to be able to answer: what did the agent see, what did it decide, why, and who is accountable? This layer records an agent's inputs, reasoning, actions, and trades so its behavior can be audited.
We think this is one of the most underbuilt parts of the stack. The opportunity is a new trust infrastructure for agents: agent identity and attestations ("Know Your Agent"), tamper-proof decision logs, verifiable performance tracking, and reputation systems. Companies like Alien*, t54 Labs, Lemma, and AgentOps are early movers.
Where value accrues
Here's our core view: in agentic trading, the agent itself is not the moat. Open-source models running on open-source harnesses mean anyone can spin up a capable trading agent over a weekend. If the agent is a commodity, the value has to accrue somewhere else.
We see three shifts driving where it goes:
Execution fees compress, but liquidity still wins. Agents will squeeze trading commissions, because they compare prices across venues instantly and have no loyalty to any single screen. But they optimize for the total cost of a trade, including spreads and price slippage, not just the fee. Agentic order flow will go to the deepest, tightest markets. Routing orders will be a shrinking business. Owning the liquidity will not.
Value moves to whoever owns the decision loop. We believe agentic trading platforms will increasingly earn money through what we call Intent-as-a-Service: turning a user's goal into a working portfolio, and keeping it on track. Users will pay for outcomes, not clicks. That "intent premium" belongs to whoever owns the decision loop, not just the order pipe.
Trust and breadth win at the platform level. If anyone can build an agent, the winning platform is the one that users and their agents trust to hold their assets, and that offers the widest range of assets to trade on the safest, most reliable rails. When users arrive with portfolio intent, a platform that only offers one asset class will lose the agent to one that offers them all.
For startups, this leads to a simple filter. We believe the most durable venture-scale outcomes will form where proprietary intelligence, distribution, or regulation creates lock-in. A thin agent wrapper around a general-purpose model with a trading API attached is unlikely to last.
Where we're investing
With that lens, Coinbase Ventures is focused on three pillars of the agentic trading stack:
Agentic trading harnesses and terminals. We're looking for products that go beyond routing orders to deliver true portfolio-building intent, with a strong user experience and proof that people stick around: growing agentic volume, assets under management, and retention.
Access to a real edge. As general-purpose reasoning becomes a commodity, the advantage shifts to whoever can give agents an edge others can't easily copy. AI research labs building market reasoning models that can reliably run multi-step financial workflows without supervision are one clear path. We're also interested in unique data providers, quantitative strategy models, and verified strategy marketplaces, that agents can tap into.
Compliance and observability. Regulatory frameworks aren't built for non-human decision-makers. We're investing in the infrastructure that closes that gap: agent identity and attestations, tamper-proof decision monitoring, full traceability of every tool an agent used, performance verification, and reputation.
When we evaluate companies, we focus on defensibility over raw capability:
An intelligence flywheel. Because the model alone isn't the moat, we look for "context loops," where proprietary data and specialized models reinforce each other so that strategies improve over time and the gap with competitors widens.
Agent-native teams. The strongest teams combine three skill sets: deep quantitative trading experience (for risk-adjusted returns), advanced AI engineering (for reliable agent behavior), and regulatory literacy (treating compliance as a product feature, not an afterthought).
Looking ahead
Agentic trading marks a real shift in how people interact with markets. As agents become easier to build, we believe lasting businesses will need advantages others can’t easily copy. The winners will be the ones that own the intelligence, earn users' trust, and give agents the safest regulated path to the broadest set of markets.
The stack is still being built, and much of it – from trading-hardened harnesses to strategy marketplaces to agent identity and audit trails – is wide open. If you're building in agentic trading, we'd love to hear from you.
Related work
AiFi: The Financial System for AI, Coinbase, Aug 2026
The Rise of Onchain AI: Agents, Apps, and Commerce, Coinbase Ventures, May 2025
Demystifying the Crypto x AI Stack, Coinbase Ventures, October 2024
Disclosures
The contents in this post — and available on any associated distribution platforms and any public Coinbase Ventures online social media accounts, platforms, and sites (collectively, “content distribution outlets”) — should not be construed as or relied upon in any manner as investment, financial, legal, tax, or other advice. You should consult your own advisers as to legal, business, tax, and other related matters concerning any investment. Investments in startups and early-stage businesses are inherently risky and highly illiquid. These types of investments are suitable only for qualified, sophisticated investors who can afford to bear the risk of losing the entirety of their investment capital.
Under no circumstances should this post or any other information provided on the Coinbase Ventures website — or on associated content distribution outlets — be construed as an advertisement, solicitation, invitation, recommendation, offer or inducement of any kind. Nor should this post or such information be construed as an offer to provide investment advisory, investment management, financial advisory, or financial management services. The information on the content distribution outlets should not be relied upon as research, investment advice or a recommendation of any kind.
Past results of Coinbase Ventures’ investments or investment strategies are not indicative of future results. Coinbase Ventures may in the future have a position or holding in any company referred to on the Coinbase Ventures website, any associated content distribution outlets, or otherwise.





